昇腾计算产业发展白皮书.pdf

1.1 AI 1.1.1 AI1.1.2 AI 1.2 AI 1.2.1 1.2.2 1.2.3 AI 1.2.4 2.1 2.2 3.1 3.2 3.2.1 3.2.2 3.2.3 3.2.4 3.2.5 3.3 3.3.1 3.3.2 3.3.3 AI 3.4 3.5 01010203030508091114172122222323242425262729323.5.1 3.5.2 3.5.3 3.5.4 AI 3.5.5 3.5.6 4.1 AI 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10 5.1 5.2 5.3 5.4 5.5 1 2 3233343536373941434547495154565759626469727979011AIArtificial IntelligenceEfficientNetTop50298.7%96%RNN-T96.8%94.17%BERT90.9%82%AIAIAIAIAIAI2019AI598.62020-202434.8%AI5G5GAIa FixEfficientNetImageNet with extra-training datab BiLSTM RNN-TThe LibriSpeech datasetc Megatron-BERTRACE03VR/AR5GAIAI+AIAI2025AI86%80%AIAIAIAIAIAIAIAIAI20126AI3004BERT,GPT-3TFLOPSPFLOPSEFLOPSAIAIAIAIAIAI05 +Fine tuningNPU API AIAI06AIAIAI07AIAIAIAIAIAIAIAIDSADomain Specific ArchitectureAIAIAI08APIAIAPIAIAPIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAI09AIAIAI5001:10AI72%AIAI0.410.080,090.130.17AIAIAIAI10112AICANNCompute Architecture for Neural NetworksAI12OEMODMIHVC&SIISVXaaSMindSporeTensorFlow/PyTorch/CANNNPUModelArtsHiAI ServiceMindXFusionDirector/SmartkitsMindStudio13AIAIAI14 AI CANNMindStudio AIMindSporeMindXModelArtsHiAIIaasPaas,SaaSAI15Atlas AIAtlasAtlasAIAIMindSporeAITensorFlowPyTorchCaffeMindSporeAIAIAIAI16173AI3-118AIMindX EdgeMindX EdgeMindX DLMindStudioFusionDirector/SmartkitsModelZooModelZooModelZooSDKSDKSDKHiAI Service/MindSporeModelArts TensorFlow/PyTorchCANNAscendCL/KernelNPU Driver19Atlas320 TFLOPS FP16Atlas256P1024P FLOPS FP16ResNet-50ImageNet25.9sAtlas88 TOPS INT880 AtlasAIPUE(Power Usage Efficiency)1.1E512 TOPS/WAI MindXModelArtsHiAIMindXSDKSDKSDKISVAIAIMindSporeMindSpore20CANNMindStudioAscendCLAscend Computing LanguageCANNAscendCLAIAIAPIAscendCLAIAIMindXAIMindX DLMindX EdgeModelArtsHiAI ServiceMIndX ModelZooSDKModelZooSDKSDKISVAIAICANNAI21MindSporeAIMindSporeAIAIAIAtlasAIAtlas 200 AIAtlas 200 DK AIAtlas 300 AIAtlas 500Atlas 800 AIAtlas 900 AIAtlasAtlasAtlas 200 AIAIAtlas 300 AIAIOEM/ODM22AINPUNeural Processing UnitNPUAI16*16*163D Cube409616MACIPAI22TOPS INT811 TFLOPS FP168WAIIT320 TFLOPS FP16SoCAICPU CoreDVPPTask SchedulerHost CPUHCCSPCIe 4.0RoCE v2Scale OutScale Up?23?Atlas 200 AI?3000?22TOPS INT8?AI?Atlas 300I?3000/3010?88 TOPS INT8?80?2?Atlas 300T?9000?320 TFLOPS FP16?3/4?PCIe?AI?Atlas 500?Atlas 800?Atlas 500 Pro?24Atlas 8003000/3010/Intel2UAI8AI704 TOPS INT8OCRAtlas 8009000/9010Intel4U82.24P FLOPS FP16Atlas 500 Pro 30002UAI4AI352 TOPS INT8AI475mm600mmAtlas 900 AIHCCS PCIe 4.0 100G RoCE256P1024P FLOPS FP1650PCAI25 TBECANNAIAICANN10EMUIAndriodopenEulerUOSUbuntuDebianSuse14AICPUNPUAICANN26CANN4Driver HCCLHuawei Collective Communication Library 100GRoCEAscendCLPluginAIAscend-IR(Intermediate Representation)CANNTBE-DSLTensor Boost Engine-Domain Specific LanguageDSL70%TBE-TIKTensor Iterator KernelMindStudioMindStudio27MindStudioMindStudioMindStudioIntelliJAIAIAIAIAIMindSporeMindSporeMindSporeAuto ParallelMindSporeSCTSource Code Transformation,ADAutomatic differentiation,SCT28MindSporeAIMindSporeMindSporeSCTADMindSporeAuto ParallelMindSporeMindSporeAIAIMindSpore129234MindSporeSDKSDK30MindX 2(MindX DLMindX Edge)1(Model Zoo)SDKKubernetesHCCL-ControllerDevice Plugin VolcanocAdvisor for AscendMindX DLMindX DLNPUAscend Device PluginHCCL-ControllerVolcanocAdvisor for AscendAIMindX EdgeAIMindX EdgeAgent31CFDHA AgentEdge SiteEdge CoreKubeEdgeEdge HubMindX EdgeMindX Edge256MCPU3%MindX EdgeAIEdge AgentdockerEdge CoreEdge SiteEdge HubEdge SiteModelZoo32ModelZooAIMindSporeTensorFlowPyTorchCaffeAISDKAIAISDKSDKSDKSDKSDK5AI0.510SDK33CPUGPU,NPU,FPGA/MindSporeMindSpore2020328https:/CANNAIMingSporeCPU+GPU+NPU+FPGA34https:/ AIImperial College of LondonUniversity of EdinburghUniversity Paris-SaclayUniversitt Mnster13/MindSporeLFAILinux TACOutreachLFAIAcumosONNXONNXOpenI-OpenIApacheTVMAdvisory BoardTOP102securitysafetyAIAI35AIDVPPDigital Vision Pre-processingAIPPAI Pre-processingAINPUshape/AIDVPPAI CoreAIPPAI Core/36CANN10003GPT-317501TBPipelineOn-DeviceHostCPU37CUBE/VECTOR/CPU/SDMAUBL1/L2SuperKernelDDB/PipelineSIMDMindSporeTuring-NLG,GPT-338Cost ModelCost Modeltensortensortensorpipeline394AI40AI(/)AIX86AIGPU/HCSO(ModelArts)Atlas 800Atlas 900 AIAtlas 800Atlas 900 AIAIAIAIAIAIAIAIAIAI1)AI41L0-L32)AIAtlasHCSOModelArtsAI50%80%70%AI4.51.83)AIMindX DLweb510IIAtlas 900 AIAIAIEAIAI42AIAIAIAtlasAIAtlas BERT-Large ResNet5080%PUE1.1Atlas808640Atlas43ResNet50BERT-Large1TensorFlow3D+Batch2AIVRAIAI44AI99%Atlas 800 AIAtlas 800 LEDAIAIPCBPrinted Circuit Board4590%99.9%33AIAI46AIAtlas 500AGVAutomated Guided VehicleAIAGVAtlas 200985AI47AIAIAIAtlas AI48Atlas 200 AIAI1/1/Atlas 200 AIAtlas 800 Atlas 500 Pro49Atlas 200 AIAIAtlas 500AI 20 28050OCRAtlas AtlasOCROCR51MindX DLAIAIAIOCRAI52 Atlas AIAtlasAtlas53Atlas AIAIISVAtlas AI1:11N2.054AIAIAI2B2C2B5G+AI2CAIAIAIVIP95%Atlas 50070%50%VIP60%55AtlasCPU/GPU/NPUAIAIAIAIAIAI5G Massive MIMOiRayceNAIEAUTINAtlas 20030%20%5G AtlasAI5GAI2BAtlas 800 AI56AI5G5G+AI2B5G+AI201920202020202057Atlas 500 Pro99%AI-AI2.5AI98%20Atlas15AIAI58AICTAtlas AIAtlas 800 AIPACSPicture Archiving and Communication SystemsCTCTAICTAI10-1510-1590%99%59560AIAI+AIAI+AIAIAIFAE 61AIAIAIAI+AIISV62AIIAAITISAGCCHAIAIHDCHCAIAIMindSporeAIAITensorFlowPyTorch,CaffeAIAI63ISVISVOpenlabISVISVISV64AI90%AI10%AI1%MindX MindSporeCANNMingStudio 2.05100AIMindXAIAIMindSporeAI65CANNAscendCLAPI100505070AI20166AtlasMindStudioMindStudioMindStudioAIBenchmarkAIMindX DLWebUIMindSporeTensorFlowPytorch AIMindStudioAIACLAI/C+APIAIAIAIDocker67SDKAIMindStudioAIAIAI50AI68AIAIHCIAHuawei Certified ICT AssociateICT/HCIPHuawei Certified ICT ProfessionalICTDevRunAIAIAIAtlas AI1220169AIAIHAGHuawei Ascend GroupAIAIAIAIAI/70AIAIAIAI2000-2020721)/2)/3)/4)712.0-AIAIAI AIMindSpore72AI Developer73LDFPDFMDFMDFNREMDFPAE/FAEMarketplaceC&SIXaaSISVOEMIHVODM NREOpenLabAtlas AI74MarketplaceAI HCIAAIAIAI75PPTMOOCAIAI/WorkshopAIVCAtlasAIAtlas 200DKMarketplaceAtlas AIAI2020100AI766SoC78AIAISDK AIAIAIAIAIAI1 Fixing The Train-Test Resolution Discrepancy:FixEfficientNet2 Improved Noisy Student Training for Automatic Speech Recognition3 Megatron-LM:Training Multi-Billion Parameter Language Models Using Model Parallelism4 GIV 2025 5 AI and Computehttps:/ AI7 ElementAI2019AI8 AI2019-2020 AI Pre-processing Ascend Computing Language Ascend Intermediate Representation Compute Architecture for Neural Networks Digital Vision Pre-processing Huawei Ascend Expert Huawei Ascend Group Huawei Collective Communication Library ICTICT Infrastructure as a Service Platform as a Service Picture Archiving and Communication Systems Peripheral Component Interconnect Express Software as a Service Tensor Boost Engine Tensor Iterator Kernel RDMA over Converged EthernetRDMA